AnAdaptation ofEPCAtoImageCompression andReconstruction
Zhili Wu · 2006
Principal Component Analysis (PCA)andother SVD related approaches arecommonly usedindimension reduction andreconstruction ofimages. However, aslinear methods theymaynotbeappropriate forsomenon-linear cases. Recently anewapproach namedasExponential Family Prin- ciple Component Analysis (E-PCA) isproposed fornon-linear compression andhasbeensuccessfully usedtosolve thebelief states' dimension reduction ofPartially observable Markov Decision Process (POMDP). Inthispaper, we attempted to adaptE-PCAtoimagecompression andreconstruction dueto thereason thatitcanguarantee nonnegative reconstruction and isfit forsomenonlinearly distributed data. Theoriginal E-PCA formulations arealsosimplified inthis papertoaccelerate the parameters learning process. Experiments areperformed on somestandard imagedatasetstoverify theeffectiveness of E-PCAonimagecompression. Fromtheexperimental results, we canconclude thatthenewadaption ofE-PCAonimage compression isparticularly effective whentheimagedata follows somekindsofdistribution.